Hallucination firewall

Large language models are probabilistic: they generate plausible text, not verified facts. In an insurance transaction, a plausible-but-wrong premium figure, a fabricated eligibility decision, or an interpolated excess would represent a regulatory breach - a potential mis-selling event if the customer acts on it, and an audit failure regardless. The hallucination firewall is the structural answer to this problem.

In the A2A protocol, the no-fabrication rule (Core Patterns Section 2.1) states: every fact an agent presents must trace to an authoritative source, whether a carrier response, an enrichment source, or a disclosed document. The agent must not invent, estimate, interpolate, or recompute any such fact. The hallucination firewall is the enforcement mechanism for this rule at the system level.

Technically, it operates as a schema validation step in the Rail. Before any quote call is made to a carrier, the assembled request is validated against the carrier's defined field rules: types, allowed values, mandatory and conditional field logic. Before any quote response is returned to the agent, the response is validated against the canonical response schema. A field that fails validation is flagged, and the agent must surface the uncertainty using the honest-absence pattern rather than fabricate a replacement value.

The hallucination firewall also underpins the FAQ and knowledge-base tool that allows insurers to ground AI responses in approved product content. Rather than an LLM generating a description of cover from training data, it pulls from an insurer-approved knowledge base, reducing hallucination at the information layer as well as the transaction layer.